Statistical analyses
Analyses of the above trait values were performed in four steps with R 3.5.2 (R core team 2018).
First, we tested for divergent trait evolution in plants descending from dry, control and wet manipulated plots in the central sites SA and M (N=240 genotypes). For each trait separately, linear mixed models were calculated with climate manipulation treatment (dry, control, wet), site (SA, M), five greenhouse watering levels, and their interactions as fixed factors, as well as genotype as random factor (accounting for five non-independent plants across water levels). Some traits were transformed prior to analyses to meet homoscedasticity (sqrt: stomata density, height, reproductive allocation, seed number; log: leaf number at flowering, vegetative biomass). Significance was assessed with Wald F-tests with Kenward-Roger approximated df in the package car (Fox & Weisberg 2011). Posthoc tests identified contrasting climate manipulations using the package ‘multcomp’ (Hothorn et al. 2008) with P-values corrected for false discovery rate (FDR) sensuBenjamini & Hochberg (1995). For germination fraction (binary) we used a corresponding glm with logit link-function and quasibinomial error structure.
Second, we tested for clinal trends in traits across the rainfall gradient, including only plants descending from control plots in all four sites (N=160 genotypes). We calculated linear mixed models per trait with site and greenhouse water level as fixed factors, and genotype as random factor (transformations as above). Posthoc tests with FDR-correction as above identified contrasting sites. Germination fraction was analyzed with a binomial glm as above, using only site as main factor.
Third, we estimated selection, i.e. the covariance of traits with relative fitness (Lande & Arnold 1983), under low and high irrigation in the greenhouse. This approach reveals traits that can adapt a population to drought and is independent of other environmental factors correlating with rainfall (Mitchell-Olds & Schmitt 2006). We estimated selection for all traits showing either rapid evolution (step 1) or clines with rainfall (step 2). We included all plants from sites with climate manipulation, computed the genotype trait-mean across low watering (15ml, 20ml) and high watering (50ml, 90ml), followed by standardization (zero mean, 1 SD) per population (SA and M) and watering level. Similarly, relative fitness was computed per population and watering. We fitted generalized least squares models (gls, rmspackage (Harrell 2019)), with relative fitness as the dependent variable, and trait, watering (high, low) and their interaction as predictors. A significant trait × watering interaction indicated contrasting directional selection on that trait contingent on water availability, computed using type III sums of squares (Anova, carpackage (Fox & Weisberg 2019)) with FDR-correction.
Fourth, we tested whether field climate manipulations favored genotypes with higher plasticity. In addition to assessing the climate manipulation × water term in step 1 above, plasticity was quantified for the above traits using the Coefficient of Variation (CV) across the five individuals (i.e. water levels) per genotype in the greenhouse. The intuitive, standardized CV allows comparing plasticity across different traits (Houle 1992) and handles well outliers and non-linear responses across several environments. Another plasticity index (PIv, see Valladares et al. 2006) yielded the same results. With these CV-values per genotype, we calculated two-way ANOVAs and FDR-post hoc tests separately for each trait, including the factors site (SA, M) and climate change treatment (dry, control, wet).